Main

The extracellular matrix is a key interface that actively regulates cell–cell interactions, host–pathogen interactions, and autocrine and paracrine signalling through growth factor binding (reviewed in refs. 13,14). HSPGs are a central component of the extracellular matrix and consist of a core protein and one or more covalently linked heparan sulfate (HS) chains1. The protein interactome of HSPGs is well established and includes chemokines, cytokines, growth factors, morphogens, cell-adhesion proteins and viral components such as the SARS-CoV-2 spike protein5,15,16,17,18. HSPGs are conceptually considered as coreceptors that facilitate the formation of ligand–receptor complexes. Protein–HS interactions occur via sulfated domains of the chains, and high-level and/or site-specific sulfation of HS control the various activities ascribed to HSPGs16, which span development, physiology and disease.

In addition to HS, RNAs represent another class of negatively charged biopolymers on the cell surface (cell surface RNAs (csRNAs))9,19,20,21. Among them are glycoRNAs9, which contain sialylated and fucosylated N-glycans on small RNAs and can be detected by the sialic acid-binding immunoglobulin-type lectins (Siglec) family of receptors9. A covalent linkage between small non-coding RNA and N-glycans has been identified as the modified RNA base 3-(3-amino-3-carboxypropyl)uridine (acp3U)10. Functionally, N-glycosylation of RNA at acp3U has been shown to chemically shield this RNA modification from sending by endosomal TLRs22. GlycoRNAs have also been found to interact with lectins (for example, P-selectin) and control neutrophil recruitment in vivo23. In addition to glycoRNAs, the surface of cells also contains a selective suite of RBPs that are in physical proximity to glycoRNAs, form precise nanosized clusters and are sites of cell-penetrating peptide (for example, TAT) entry, regulated by RNA11. These csRBPs have been demonstrated on many different cell types and certain csRBPs can be selectively presented on cancer cells, in the cases of csNPM1 (ref. 12) and csU5 snRNP200 (ref. 24). Together, these reports highlight that csRNAs are not merely discarded passenger molecules but important biological players in cell surface biology. However, the specific roles of these csRNP clusters in cell signalling and development remain poorly understood.

Although insights into csRNPs are emerging, extracellular RNA (exRNA) and extracellular RNases have long been characterized, there is long-standing evidence that exRNA biology may contribute to angiogenesis. It is thought to modulate VEGF signalling; however, the precise mechanisms by which exRNA and its associated RNPs exert this regulatory effect remain unclear. For example, exRNA has been proposed to modulate VEGF activity by directly interacting with both the C terminus25 and the N terminus26 of the VEGF-A165 isoform. This RNA-mediated process appears to be sensitive to RNase treatment, as it can abrogate the observed VEGF stimulation25,27. Separately, angiogenin (ANG), an extracellular RNase, is known to promote angiogenesis (reviewed in ref. 28), yet whether it targets exRNAs is poorly defined. Of note, pharmacological inhibition of ANG orthologues in zebrafish broadly impairs vascular development, suggesting that exRNA degradation itself may have an essential regulatory role in angiogenesis during development29. Thus, despite decades of research generally on angiogenesis30,31 and RNase-mediated or exRNA-mediated angiogenic pathways27,28, it remains unresolved which specific RNA species are involved, through what receptors or molecular partners they act, and the exact mechanisms on the cell surface by which they modulate angiogenesis.

Here we uncovered a new layer of regulation of csRNPs as potential antagonist of HS-mediated signalling. We found that csRNP assembly on living cells depends on intact HS chains modified with N-sulfate and 6-O-sulfate. Functionally, VEGF-A165 can directly bind to RNA on the surface of endothelial cells, which negatively regulates its intracellular signalling cascade. The heparan-binding C terminus of VEGF-A165 can bind to glycoRNAs. Mutating key arginine residues to lysine in this domain decouples HS binding from glycoRNA binding and enabled on cell and in vivo experiments, revealing hyper-vascularization of the mouse retina and the developing vasculature in zebrafish embryos in this mutant. Together, our work nominates csRNPs as an antagonist of VEGF-A165-mediated signal transduction in endothelial cells and during angiogenesis.

Siglec-11 binds to csRNA

We previously found that Siglec-11, a member of the Siglec receptor family32, which are cell surface receptors, binds to the surface of HeLa cells in an RNA-dependent manner9, suggesting that Siglec-11 is potentially a csRNA receptor. To explore whether other members of the Siglec family are also RNA-dependent cell surface binders, we used confocal imaging and screened the 13 commercially available Siglec–Fc fusion proteins for binding to suspension (MOLM-13) and adherent (U2OS) cell lines. Only Siglec-4, Siglec-7, Siglec-9 and Siglec-11 specifically bound both cell types and only Siglec-11 was reduced after live-cell RNase treatment11 (RNase A and RNase III; Extended Data Fig. 1a–c). By contrast, although Siglec-7 and bulk cell surface sialic acids (periodate labelling33) were lost after live-cell sialidase treatment, Siglec-11 binding was not impacted (MOLM-13) or only mildly impacted (U2OS; Extended Data Fig. 1d). To test whether Siglec-11 was indeed binding near RNA, we used an anti-double-stranded RNA (dsRNA) antibody (9D5) that detects csRNA11. Co-staining of U2OS and MOLM-13 cells with 9D5 and Siglec-11 showed that 42% and 61% of 9D5 puncta overlapped with Siglec-11 on MOLM-13 and U2OS cells, respectively (Fig. 1a and Extended Data Fig. 2a,b), 34.5% and 24% of Siglec-11 puncta overlapped with 9D5 on MOLM-13 and U2OS cells, respectively. Co-staining 9D5 with Siglec-7 and separately with Siglec-9, both of which are not sensitive to RNases, showed that only 2.5% and 1.4% of 9D5 puncta overlapped with Siglec-7 or Siglec-9 in MOLM-13, and 0% and 2.8% of 9D5 puncta overlapped with Siglec-7 or Siglec-9 in U2OS, respectively. Finally, although the cell surface binding of Siglec-11 is sensitive to RNase, we next assessed whether Siglec-11 can directly interact with sialic acid-modified glycoRNAs. Using highly purified RNA (Methods), we evaluated enriched material from immunoprecipitations with Siglec-11 or an Fc control and chemically labelled for glycoRNAs (rPAL10). We found that Siglec-11 captured sialoglycoRNAs above the Fc and that this signal was sensitive to RNase treatment (Extended Data Fig. 2c). The well-correlated binding of 9D5 and Siglec-11 suggests that Siglec-11 ligands are near glycoRNA on the cell surface.

Fig. 1: csRNA and RBP clusters are dependent on HS biogenesis.

a, Representative images of U2OS cells co-stained with Siglec-11 (yellow) and 9D5 (magenta). An enlargement of the hatched box is shown, and a bright field (BF) is shown to represent the outline of the cell membrane. The nearest neighbour distance analysis of the Siglec pairs is also imaged (bottom). For each pair, the distance (µm) from that anchor (left side protein name in the figure key) to the other pair was calculated across the indicated number of cells. These values were plotted in a density histogram. b, Dot plot of genes identified in the genome-wide CRISPR-KO screen for the loss of Siglec-11 cell surface binding ranked by CRISPR score. The top 15 gene names are displayed, with cut-off shown. The inset illustrates how Siglec-11 could interact with a glycoRNA. c, Dot plot of genes identified in the genome-wide KO screen as in panel b, for the loss of 9D5 cell surface binding. The inset illustrates how 9D5 could interact with a glycoRNA. d, Upset plot analysis of genes with a score cut-off of −0.8 from the Siglec-11, 9D5 and MAA-I genome-wide KO screens. The common overlapping hits between 9D5 and Siglec-11 are highlighted in blue. The total number of hits for each intersection is noted. Statistical assessment was performed with a hypergeometric test and P values are shown without adjustment. e, Representative images of WT and KO U2OS cells stained live with 10E4, Siglec-11–Fc, 9D5 and Siglec-7 (all in red). Ab, antibody; dsRNA, double-stranded RNA; mE, mEmerald. f, Quantification of the indicated signal dot numbers and intensity per cell from panels e,g from three independent experiments. Mut, mutant. g, Representative images of the indicated cell lines stained live with anti-DDX21 and anti-hnRNP-U (both in red).

HSs organize csRNPs

To identify what mediates Siglec-11 and 9D5 binding, a genome-wide CRISPR–Cas9 gene-knockout (KO) approach34 was developed based on flow cytometry of MOLM-13 cells using Siglec-11, 9D5 and the plant lectin MAA-I, which is well characterized to selectively bind to sialic acid35 (Fig. 1b,c and Extended Data Fig. 3a). We identified 154, 187 and 246 hits in the Siglec-11, 9D5 and MAA-I screens, respectively (Fig. 1d and Supplementary Table 1). The top hits enriched in the MAA-I screen were related to sialic acid biosynthesis (for example, CMAS (#1) and NANS (#11)36; Supplementary Table 1), confirming the specificity of the screen; however, both of these genes scored poorly in the Siglec-11 and 9D5 screen. Intersectional and pairwise analysis of these enriched genes showed a more robust overlap between 9D5 and Siglec-11 than MAA-I with either of the other two probes (Fig. 1d), consistent with Siglec-11 association to the cell surface in an RNA-dependent manner. Direct inspection of the 41 genes commonly scoring between 9D5 and Siglec-11 showed consistent robust scores for each hit, whereas these genes often scored poorly in the MAA-I screen (Extended Data Fig. 3b), highlighting the overlap and common effect of these genes on both 9D5 and Siglec-11 cell surface binding. Gene ontology analysis revealed terms enriched in membrane compartments and processes including heparin biosynthetic process (Extended Data Fig. 3c). Inspection of the top hits of Siglec-11 revealed EXT1, EXT2 and UXS1 as the top three genes, which are key enzymes of HS biogenesis (Fig. 1b). The 9D5 screen also revealed EXT1 as a highly scoring hit, whereas UXS1 and EXT2 were present but below the cut-off.

To verify that HS chains are required for the tethering of Siglec-11 and RNA to the cell surface, we generated KO clones in U2OS cells. Within the Golgi, HS chains are initiated by xylosylation of proteoglycan core proteins, which depends on UDP-xylose formation catalysed by UXS1 (refs. 1,37). This is subsequently elongated by a hetero-dimeric complex formed by EXT1 and EXT2 by adding alternating residues of D-glucuronic acid (GlcA) and GlcNAc to the nascent polymer38,39,40, providing the eventual substrate for sulfation and HS-interacting proteins to bind. We stained KO cells (Extended Data Fig. 4a,b) with 10E4, a specific antibody recognizing mature HS41,42, and verified that HS was not produced (Fig. 1e,f). Moreover, EXT1 KO,* EXT2* KO and UXS1 KO cells all show complete loss of both Siglec-11 and 9D5 binding. The binding of Siglec-7 and Siglec-9 was not affected on EXT2 KO cells (Fig. 1e,f and Extended Data Fig. 4c,d). The loss of 9D5 signal was selective to the cell surface as intracellular staining revealed no change inside of cells upon EXT2 KO (Extended Data Fig. 4e). To understand whether this effect was due to the enzymatic activity of EXT2, EXT2 KO cells were transfected with wild-type (WT) EXT2 or a catalytically inactive (D517N/D573N40) mutant cDNA (Extended Data Fig. 4f). Reintroduction of WT EXT2 rescued the loss of 9D5 and Siglec-11, whereas the catalytically inactive mutant did not (Fig. 1e–g). These effects were not restricted to *EXT2 *KO, as *EXT1 *KO and *UXS1 *KO also resulted in the loss of HS, 9D5 and Siglec-11 binding (Extended Data Fig. 4g–i). These results show the dependency of clustered csRNA on HS biosynthesis.

Given the colocalization of glycoRNA and csRBPs11, we next tested whether csRBPs are also organized by HS. On WT U2OS cells, both cs-DDX21 and cs-hnRNP-U colocalize with 9D5 (Extended Data Fig. 4j). Consistently, no observable DDX21 and hnRNP-U antibody binding occurred on the cell surface in EXT2 KO cells (Fig. 1f,g). This loss was not due to bulk changes in DDX21 or hnRNP-U as whole-cell lysates had no changes between the WT and KO (Extended Data Fig. 4k). The formation of csRBP clusters is also dependent on EXT2 enzymatic activities as reintroduction of the WT, but not the catalytically inactive (D517N/D573N)40, rescued the formation of csRBP clusters in EXT2 KO cells (Fig. 1f,g). These findings indicate that HS is a critical scaffold in modulating the recruitment and retention of RNA and their respective binding proteins (for example, csRBP), csRNPs, to the cell surface.

csRNPs depend on intact HS polymers

To determine whether intact HS chains on the cell surface facilitate csRNP clustering, we performed cell staining on U2OS cells after heparin lyase treatment (30 min, heparin lyases I/II/III). Treated cells lost the binding of 10E4, Siglec-11, 9D5, anti-DDX21 and anti-hnRNP-U, whereas Siglec-7 and Siglec-9 binding was not impacted (Extended Data Fig. 5a,b). Given the colocalization of 9D5 and Siglec-11 and sensitivity to heparin lyases, we sought to understand the spatial and temporal relationship between 9D5, Siglec-11 and HS. We therefore performed a three-colour co-staining experiment on U2OS cells and found that Siglec-11 puncta are highly correlated in localization with 10E4 puncta (Fig. 2a and Extended Data Fig. 5c,d, green line). This relationship was also true when examining 10E4, Siglec-11 and anti-DDX21 (Extended Data Fig. 5e), supporting the clustering model of csRNPs near to sites of HSPGs. We next assessed the temporal regulation on csRNP clustering by conducting a recovery experiment in cells after heparin lyase treatment, removal of the enzymes and incubation of the cells with fresh media for 0, 45, 90 or 180 min to access de novo HS production and cell surface presentation (Fig. 2a and Extended Data Fig. 5c,d). HS puncta reappeared 45 min after a recovery from heparinase treatment (9.4%). At this early time point, Siglec-11 and 9D5 clustering was not detected. After 90 min, however, 9D5 and Siglec-11 started to recover together at sites of large HS clusters, suggesting that a critical concentration of HS is needed for proper csRNP clustering. After 180 min, the levels of HS, Siglec-11 and 9D5 binding recovered to native levels, confirming that intact HS chains are vital for csRNP cluster formation on the live cell surface.

Fig. 2: csRNPs are dependent on intact HS chains.

a, Representative images of U2OS cells treated with a heparinase pool for 30 min, recovering for the indicated times and co-stained with 10E4 (cyan), Siglec-11 (yellow) and 9D5 (magenta). An enlargement of the hatched box is shown, and a bright field is shown to represent the outline of the cell membrane. Three independent experiments were performed. b, Schematic of a HSPG with the regions of activity for the various enzymes and HS oligos perturbed in the following experiments. 2S, 2-O-sulfation; 6S, 6-O-sulfation; NS, N-sulfation. c, Representative images of WT, NDST1-KO-treated, HS6ST1-KO-treated, HS2ST1-KO-treated, Sulf1 stably overexpressing (OE)-treated, Sulf2 stably OE-treated, Tega HS #09-treated and Tega HS #37-treated U2OS cells stained with 10E4, Siglec-11, 9D5, anti-DDX21 and anti-hnRNP-U (all in red) separately. d, Quantification of data in panel c with the number of cells counted from three independent experiments. Data are mean ± s.e.m.

6-O-sulfated HS facilitates csRNPs

The complex but precise sulfation of HS chains16 is initiated by one or more members of the NDST family of N-deacetylases-N-sulfotransferases (NDSTs) acting on a subset of GlcNAc residues (Fig. 2b). NDST1 KO in U2OS cells (Extended Data Fig. 6a–c) resulted in a 74% loss of N-sulfoglucosamine residues as measured by reduction of 10E4 staining (Fig. 2c,d). NDST1 deficiency caused a 77%, 72%, 62% and 63% loss of binding intensity per cell in Siglec-11, 9D5, cs-DDX21 and cs-hnRNP-U, respectively (Fig. 2c,d), suggesting that sulfation of HS chains is required for tethering RNA and associated proteins to the cell surface. To broadly and non-genetically inhibit sulfation, we treated cells with sodium chlorate, a metabolic inhibitor of sulfation43, and found near complete loss of 10E4, Siglec-11, 9D5, cs-DDX21 and cs-hnRNP-U, whereas sodium chloride (control) had no effect (Extended Data Fig. 6d). Correspondingly, 3G10 antibody specific for cut HS41, only bound cells after heparin lyase treatment (Extended Data Fig. 6e).

Following N-sulfation, partial epimerization of adjacent GlcA residues to L-iduronic acid (IdoA) can lead to 2-O-sulfation (catalysed by Hs2st), 6-O-sulfation of GlcNAc and GlcNS residues (by HS6ST1–3)37,44. We therefore generated HS2ST1-KO and HS6ST1-KO U2OS cells (Extended Data Fig. 6a–c). Loss of 2-O-sulfation via HS2ST1 KO did not significantly alter the binding of the antibody panel, whereas loss of 6-O-sulfation through HS6ST1 KO reduced Siglec-11, 9D5, cs-DDX21 and cs-hnRNP-U binding by 76%, 67%, 56% and 53%, respectively (Fig. 2c,d). To complement this result, we individually overexpressed Sulf1 or Sulf2, two extracellular sulfatases that can remove 6-O-sulfation from intact HS45,46 (Fig. 2b and Extended Data Fig. 6f,g). Indeed, overexpression of these sulfatases removed total Siglec-11, 9D5 binding, cs-DDX21 and cs-hnRNP-U on U2OS cells (Fig. 2c,d). Finally, we performed a competition experiment by adding exogenous HS chains with high N-sulfation, 6-O-sulfation and 2-O-sulfation (rHS09) to the media. rHS09 caused a loss of clustering of Siglec-11, 9D5, cs-DDX21 and cs-hnRNP-U, whereas the addition of HS chains with only high N-sulfation and 6-O-sulfation (rHS37) increased the average level of binding twofold to threefold (Fig. 2b–d). Although both rHS37 and rHS09 are highly sulfated, rHS37 merely lacks the 2-O-sulfate groups. Combined with the genetic evidence, this indicates that N-sulfation and 6-O-sulfation of HS promotes csRNP formation.

csRNAs regulate VEGF-A165 signalling

Functionally, HSPGs bind to extracellular growth factors, facilitating signal transduction by forming a growth factor–growth factor receptor complex on the cell surface16. VEGF-A is a prototypical pro-angiogenesis factor causing endothelial cell migration and proliferation, with certain forms dependent on 6-O-sulfated HS for cell binding7,47. Upon activation by VEGF-A, the VEGF receptor (VEGFR), a receptor tyrosine kinases expressed on endothelial cells48,49, signals through activation of RAS–RAF1–MEK, which in turn phosphorylates ERK1/215. Given this, we tested the presence and enzyme sensitivity of csRNPs on human umbilical vein endothelial cells (HUVECs). HUVECs generally stain positively for the same ligands tested on cancer cells and heparin lyase treatment resulted in a complete loss of binding across the antibody panel (Fig. 3a and Extended Data Fig. 7a). Treatment of HUVECs with RNases reduced binding of Siglec-11, 9D5, cs-DDX21 and cs-hnRNP-U puncta by 73%, 83%, 96% and 93%, respectively, without affecting 10E4 staining. These features of csRNP assembly, including heparin lyase and RNase sensitive, were also consistent on primary cultures of keratinocytes (Extended Data Fig. 7b,c), confirming that control of csRNP clustering by HSPGs is conserved on primary cells such as endothelial cells and keratinocytes.

Fig. 3: Intact csRNPs repress activity of VEGF-A165 on endothelial cells.

a, Quantification of 10E4, Siglec-11, 9D5, cs-DDX21 and cs-hnRNP-U intensity per cell from three independent staining experiments on HUVECs. Data are mean ± s.e.m. a.u., arbitrary units. b, Western blot analysis of whole-cell lysates isolated from HUVECs after starvation and treatment with an RNase pool followed by 3 ng ml−1 VEGF-A165, VEGF-A121 or EGF stimulation (top). Quantification of the ratio of pERK to total ERK was also calculated across the biological triplicates (bottom). Statistical assessment was performed with a two-sided Student’s t-test and P values are shown. Data are mean ± s.e.m. c, Representative images of starved HUVECs treated with an RNase pool, and subsequently with VEGF-A165 or VEGF-A121, finally stained with anti-VEGF-A165 (red) or anti-VEGF-A (red). Statistical assessment was performed with a Student’s t-test and P values are shown. Data are mean ± s.e.m. NS, not significant. d, Representative images of HUVECs treated with an RNase pool and stained with anti-VEGFR2 (red). Statistical assessment was performed with a Student’s t-test and P values are shown. Data are mean ± s.e.m. e, Representative images of starved HUVECs treated with VEGF-A165 and then co-stained with anti-VEGF-A165 (purple) and Siglec-11 (yellow). Three independent experiments were performed. f, Schematic of the microfluidic chip (top left) used to grow HUVECs without (top middle) and with (top right) RNase A for 6 days. Representative images of BFP expressed in the HUVECs. Statistical assessment of the total migration area was performed using an unpaired two-sided Student’s t-test (bottom). Data are mean ± s.e.m. Four independent experiments were performed. g, Representative image (maximum z-projection view (left) and z-projection slice view (right)) of sprouts from a +RNase A device for the cells (blue), F-actin (red) and PECAM1 (green). Four independent experiments were performed.

Various proteoforms of VEGF-A, including VEGF-A121 and VEGF-A165, bind to the VEGFR via their N-terminal domains; however, VEGF-A165 contains an extended C terminus that enables interactions with HS and neuropilin-1 (a VEGFR coreceptor)50,51. We hypothesized that this domain could also interact with glycoRNAs. To test this, we serum starved HUVECs and used western blotting to detect total and phosphorylated ERK (pERK) after VEGF-A stimulation. Upon stimulation with 3 ng ml−1 of VEGF-A165 or VEGF-A121 for 5 min, we observed an increase in pERK levels in serum-starved HUVECs (Fig. 3b, brown bars). Pre-treatment of HUVECs with RNases before the addition of VEGF-A165 resulted in a 3-fold and 2.1-fold increase in pERK levels compared with cells without RNase treatment (Fig. 3b, green bars), and VEGF-A121 activity was RNase insensitive. The RNase-dependent effect of VEGF-A165 (and the insensitivity of VEGF-A121) is also seen at 25 ng ml−1 (Extended Data Fig. 8a–c). The RNase-dependent enhancement occurred through VEGFR2 signalling (at tyrosine 1175), as we observed detectable pVEGFR2 that was enhanced by 2.69-fold upon RNase treatment of the HUVECs (Extended Data Fig. 8d). Consistently, in vitro binding analysis with microscale thermophoresis showed that only VEGF-A165, but not VEGF-A121, was capable of interacting with small RNA (Extended Data Fig. 8e). These findings suggest that csRNA antagonizes VEGF-A165 activation of ERK signalling at least in part through VEGFR2. This effect was specific to VEGF-A165 as epidermal growth factor (EGF), which also signals through the ERK pathway but does not have a HS-binding domain52, was able to fully induce pERK in the presence of RNase (Fig. 3b). These data suggest that growth factors bearing a HS-binding domain are probably subject to RNase modulation.

We next assessed whether VEGF-A165 is recruited to csRNP clusters and whether they impact cell surface association of VEGF-A165. Serum starvation and pre-treatment with the RNases led to approximately two times more binding of VEGF-A165 to the cell surface of HUVECs, whereas VEGF-A121 was unchanged (Fig. 3c). The observed increase in VEGF-A165 cell binding was not due to changes in its receptor VEGFR2 (ref. 3), as we found no changes in the receptor abundance on the cell surface after RNase treatment (Fig. 3d). Co-staining of VEGF-A165 and Siglec-11 on HUVECs revealed 78% of Siglec-11 puncta overlapped with VEGF-A165 and 40% of VEGF-A165 puncta overlapped with Siglec-11 (Fig. 3e and Extended Data Fig. 8f,g). These data support the binding of VEGF-A165 at assembled csRNPs and that RNase treatment selectively increases VEGF-A165 cell binding.

RNase-dependent changes in pERK occur quickly (within 5 min) after VEGF-A165 addition, suggesting that downstream gene expression and cell activation could be impacted in the HUVECs. To test this, we leveraged a multi-channel microfluidic chip to assess the effect of long-term RNase A treatment on HUVECs. Blue fluorescent protein (BFP)-labelled HUVECs were pre-treated with or without RNase A and loaded into the left channel with polymerized collagen in the central channel (Fig. 3f, left). Growth into the cell-free collagen region was evaluated after 6 days (fresh RNase A added daily). RNase A treatment significantly enhanced the total migration area (BFP signal) of the HUVECs into the 3D matrix (Fig. 3f and Extended Data Fig. 8h). Staining the extended HUVEC protrusions for F-actin (red) and PECAM1 (green) demonstrated cytoskeleton and cell–cell junctions, respectively, highlighting productive expansion and growth into the collagen (Fig. 3g and Extended Data Fig. 8i). 3D reconstructions of the finger-like protrusions revealed hollow, lumen-like, tubular formations, providing evidence that the RNase A-treated HUVECs can generate structures reminiscent of vessels. The enhanced growth with RNase A treatment observed in 3D further suggests that csRNA is antagonistic to VEGF-A activity.

VEGF-A165 selects RNA using arginines

Although previous reports have suggested that exRNA could interact with either the C terminus25 or N terzminus26 of VEGF-A165, our data (Fig. 3) point to the C-terminal HS-binding domain (Fig. 4a). To test whether VEGF-A165 directly interacts with RNA on the surface of cells, we starved HUVECs for 4 h, added VEGF-A165 and UV-C (254 nm) crosslinked the culture to produce covalent bonds between any RNA and VEGF-A165. We then immunoprecipitated the VEGF-A165 with a validated anti-VEGF-A antibody and found that without UV-C, no RNA was crosslinked to VEGF-A165; however, upon crosslinking, the native migrating band disappears (Extended Data Fig. 9a,b). We predicted that this was due to covalent UV crosslinking of VEGF-A165 to RNA; consistent with this, RNase treatment of the lysate after UV exposure restored the native migrating VEGF-A165 band (Extended Data Fig. 9b). Next, we performed RNA immunoprecipitation sequencing using VEGF-A165 to determine the specificity of RNA binding. Using small RNAs from HUVECs and capturing with the anti-VEGF-A antibody, we found that VEGF-A165 enriched 235 RNAs over the input and a control omitting the VEGF-A165 (Extended Data Fig. 9c,d and Supplementary Table 2). Small nuclear RNA, rRNA, small nucleolar RNA and tRNAs were most abundant in the VEGF-A165 immunoprecipitate (Extended Data Fig. 9e). Comparing these hits to the initial set of glycoRNAs (defined in HeLa cells), we found significant overlap of 34 RNAs (Extended Data Fig. 9f and Supplementary Table 2). These data demonstrate specific preference for VEGF-A165 binding to certain RNAs, which overlap with known glycoRNA transcripts.

Fig. 4: In vivo control of angiogenesis by RNA-binding defective VEGF-A165.

a, Protein structure of VEGF-A165 (left), and the amino acid sequence of VEGF-A165 WT (top right) and the arginine-to-lysine mutant (bottom right). b, Western blot analysis of whole-cell lysates isolated from HUVECs after starvation and stimulated by 25 ng ml−1 VEGF-A165HS WT or VEGF-A165HS(R/K) (left). Quantification of the ratio of pERK to total ERK was also calculated across the biological triplicates (right). Statistical assessment was performed with a two-sided Student’s t-test and P values are shown. Data are mean ± s.e.m. c, Western blot analysis of whole-cell lysates isolated from HUVECs after starvation and stimulated by 25 ng ml−1 VEGF-A165HS WT or VEGF-A165HS(R/K) with or without RNase treatment (top). Quantification of the ratio of pERK to total ERK was also calculated across the biological triplicates (bottom). Statistical assessment was performed with a two-sided Student’s t-test and P values are shown. Data are mean ± s.e.m. d, Representative images of postnatal day 6 retinas of mice 4 h post-intraocular administration of human VEGF-A165HS WT (100 ng per retina), VEGF-A165HS(R/K) (100 ng per retina) or vehicle (PBS) were immunostained with anti-ERG (red) and IB4 (white). e, Quantification of ERG numbers (left) and IB4 intensity (right) in the vascular front regions (blue dashed line area in panel d). Statistical analysis was performed using an unpaired two-sided Student’s t-test and P values are shown without adjustment. Data are mean ± s.e.m. f, Representative images of fli1a–eGFP zebrafish larvae at 36 h post-fertilization with the indicated injection of mRNA. g, Quantification of the normalized vessel length in panel f was calculated across three independent experiments. The boxplots show the median, the 25th and 75th percentiles (bounds of the box), and minimum and maximum values (whiskers). Statistical assessment was performed with a two-sided Student’s t-test and P values are shown. h, Schematic of HS–RNP complex regulation of VEGF-A signalling and angiogenesis. The schematic was created in BioRender. Flynn, R. (2025) [https: